Continuous DNA methylation deconvolution-based surrogate for B-cell differentiation state in chronic lymphocytic leukemia Authors: Jeffrey Hage, Lauren Wainman, Fiyinfoluwa Kolawole, Fred Kolling IV, Pauline Patzke, Daniel Schoolcraft, Sarah Spracklin, Prabhjot Kaur, Brock Christensen Published: 2026 ABSTRACT AbstractBackgroundChronic lymphocytic leukemia (CLL) is clinically divided categorically into IGHV mutated (M-CLL) and IGHV unmutated (U-CLL) subtypes. We aimed to generate a continuous metric of CLL epigenetic state to correspond to the many steps of B-cell differentiation from which CLL neoplasms may arise.MethodsWe measured genome-scale DNA methylation in purified CLL samples (n = 89) and utilized reference-based cell deconvolution techniques to develop a continuous metric of epigenetic similarity across a B-naive-like to B-memory-like scale (B-Index). We used B-Index to find epigenome-wide changes across CLL differentiation states. We also analyzed the epigenetic similarity of U-CLL to normal B-cell types, and epigenetic correlates of tumor burden.ResultsWe find that B-Index accurately classifies CLL into clinical subtypes (98.8%), has a stronger epigenetic signal than IGHV gene percent identity, and demonstrates additional epigenetic signal within clinical subgroups. We demonstrate that U-CLL is epigenetically more similar to B-memory than B-naive cells and reconcile previous reports of a B-naive-like epigenetic signal. The B-memory-like program of U-CLL is enriched for binding sites of transcription factors related to the germinal center activation pathway. We also identified an epigenetic signal associated with tumor burden, which was enriched for binding sites of Epstein-Barr-Virus transcription factors.ConclusionsOur analyses show that more epigenetic states are present among CLL neoplasms than are captured by the clinical classification system. Our approach can be used to generate continuous metrics of epigenetic states for other neoplasms.Subject terms: Cancer epigenetics, Computational biology and bioinformatics, Chronic lymphocytic leukaemia Introduction Chronic lymphocytic leukemia (CLL) is a mature B-cell malignancy and the most frequent leukemia in the United States, occurring primarily in older adults1. CLL is clinically separated into two subgroups based on the mutational status of the immunoglobulin heavy-chain variable region (IGHV)2,3. Cases with ≤98% identity to germline are deemed IGHV-mutated (M-CLL), indicating that the B-cell progenitor of the neoplasm has undergone somatic hypermutation; cases with >98% identity are called IGHV-unmutated (U-CLL)4. This distinction has been validated as clinically significant for predicting risk of progression3,5 and response to therapies6,7, and is concordant with multiple methods of molecular classification8–10. The cells of origin represented by U-CLL and M-CLL have been debated for over two decades11,12. For U-CLL, the DNA-methylation profile has been previously characterized as B-naive-like10,13; however, other methods of DNA-methylation analyses report larger B-memory-like epigenetic signatures14. The cause of these discrepant characterizations remains unresolved, as well as the specific features of U-CLL that correspond to each normal cell reference. Despite two tumor subtypes defined by IGHV mutational status, there exists a continuum of molecular states in the complex B-cell differentiation process that may represent different cell types of origin for CLL14. Previous work has used epigenetic signals to classify CLL into three epigenetic states10,15, or to assign CLL cases to their most similar cell of origin based on multiple references16. A continuous metric of differentiation for CLL based on the normal differentiation program has yet to be epigenetically explored and validated. A key advancement in the analysis of bulk epigenetic data has been cell type deconvolution, which takes advantage of the distinct methylation signatures of purified cell types at specific CpG loci17. Recently, peripheral blood deconvolution libraries have been expanded to resolve 12 immune cell types18, including distinguishing between B-naive and B-memory cells. Deconvolution evaluates proportions within a mixture using constrained projection and a reference matrix of purified cell types17. While this approach was developed to quantify mixtures of discrete cell types, it has yet to be utilized for predicting intermediate states in purified abnormal cell populations. Here, we develop a continuous metric for epigenetic B-cell differentiation state of CLL (B-Index) by applying DNA methylation-based cell type deconvolution methods to a clonal cell population. B-Index is able to discriminate accurately between U-CLL and M-CLL, provides additional epigenetic information within subtypes, and carries a stronger epigenetic signal than IGHV mutational status. Our normal reference-based approach also reveals that while U-CLL is more B-naive-like than M-CLL, it is more epigenetically similar to B-memory cells overall. These B-memory commonalities are enriched for many transcription factor binding sites related to B-cell activation, providing evidence for a post-activation cell of origin model for U-CLL. We also demonstrate that CLL has distinct epigenetic correlates to tumor burden. Methods Study population, sample collection, and clinical information The study was conducted under IRB protocol STUDY02001881, reviewed by the Dartmouth-Hitchcock IRB, and determined to be exempt under Category 4 for secondary research on data or specimens. Eighty-nine patients with CLL included in the study between March 2023 and March 2024 were identified by hematopathologists during routine diagnostic review of clinical specimens. Residual discarded specimens were collected for the study, and relevant clinical and biological data were obtained through chart review, including IGHV mutational status, IGHV gene percent identity, and clonal B-cell count at diagnosis. Patient information was de-identified through an honest broker system, including removal of names, geographic information, dates of birth, Social Security numbers, and other identifiers specified in the protocol. All research involving human participants was performed in accordance with relevant ethical and regulatory requirements. IGHV percent identity was identifiable in 46 cases. Patients marked as treated received treatments for CLL prior to sample collection. Any treatments for CLL counted toward classification as ’treated’, including anti-CD20 antibodies, BTK and BCL-2 inhibitors, and chemoimmunotherapy. Total lymphocyte count in the sample was determined by the CBC corresponding to the sample draw date. Tumor isolation and methylation measurement After residual peripheral blood was obtained from study participants, CLL cells were purified using immunomagnetic negative selection, with the EasySep(TM) Direct Human B-CLL Cell Isolation kit (Stemcell, Cat #19664), as per the kit protocol. The kit removes cells expressing CD2, CD3, CD14, CD16, CD56, CD61, CD66b. DNA was then extracted from the resulting purified cells, using the EZ1 Advanced XL DNA extraction platform (Qiagen, cat #9001875), using the blood card. DNA underwent Bisulfite conversion using Zymo EZ DNA Methylation Kit, and was measured using Illumina EPICv2 array-based methylation platform, according to manufacturer protocol. Methylation preprocessing and quality control Methylation data were preprocessed using the minfi (version 1.52.1)19, Sesame (version 1.24.0)20, and ENmix (version 1.42.2)21 packages in R. Sex was predicted using the median raw probe intensities on the X and Y chromosomes, using the minfi getSex function. Predicted sex was matched to labeled sex to identify any potentially mislabeled epigenetic data. Low-quality probes, meaning probes that had detection p-values > 10−6 or less than three beads available in more than 5% of samples, were identified using ENmix’s QCinfo function. All samples had less than 5% of their probes marked as low quality and were within three standard deviations of the mean bisulfite conversion intensity, and were thus retained. Beta values were generated through the Normal-Exponential-Out-of-Band (NOOB) preprocessing method using minfi, in order to account for dye-bias. Probes annotated to the X and Y chromosomes were filtered out, based on Illumina’s 20a1.hg38 annotation file. SNP and cross-hybridizing probes were filtered out using the EPICv2 mask general file from zwdzwd.github.io/InfiniumAnnotation. Probes not corresponding to CpGs, or not included in the Illumina annotation file, were also filtered. Technical replicate probe beta values were collapsed to their averages using the betasCollapseToPfx function from the Sesame R package. Beta distribution graphs were generated by Gaussian kernel density estimation, using Sheather-Jones bandwidth selector, and 512 bins. After probe filtering and collapsing replicates, 843,782 CpGs remained in the dataset. M-values, which were used for epigenome-wide association studies (EWAS) in order to better fulfill homoscedasticity assumptions in linear modeling, were calculated using the following equation: DNA methylation-based deconvolution DNA methylation-based cell type deconvolution was conducted using a 12-cell reference library18. The library contains reference beta values for 12 cell types (basophils, B-memory, B-naive, CD4-memory, CD4-naive, CD8-memory, CD8-naive, eosinophils, monocytes, neutrophils, natural killers, and T-regulators), at 1200 CpGs selected through the Identifying Optimal Libraries (IDOL) algorithm22. The library was subsetted to CpGs that overlapped with those available in our preprocessed data. Constrained projection, as implemented in the function projectCellType_CP in the FlowSorted.BloodExtended.EPIC package (version 0.99.0)23 was applied on the reference library matrix and the CLL methylation data to produce predicted proportions. Predicted proportions were normalized to sum to one per sample. Weighted reference mixture modeling To determine if the deconvolution results corresponded better to a mixture of two distinct cell populations, or a single population with individual CpG methylation states of two references (B-memory and B-naive), we created custom references for each mixed sample at each CpG. The samples were divided into tertiles based on B-memory proportion, and the tertile with the smallest B-memory proportions (n = 30, B-memory < 0.78) was tested. The top 100 most differentially methylated CpGs between the B-naive and B-memory references available in our data were used. For each mixed sample and CpG, we estimated the expected beta value as the weighted average of reference betas, using the sample’s predicted cell-type proportions as weights. This yields the beta value expected if the sample is a true mixture rather than a single population with distinct epigenetic features of B-memory and B-naive cells. For a single sample, the following equation was used to generate the sample’s mixed reference. β~i represents the predicted beta value at CpG i for a sample given the mixture of cell types, K is the number of reference cell types, pk is the predicted proportion for cell type k. βi,kref is the reference beta value at CpG i for cell type k. i is the index for each of the 100 CpG sites, and k is the index for K reference cell types. The actual beta value for each sample and CpG was then compared to the beta value of the predicted mixture, the purified B-memory reference, and the purified B-naive reference, assessing which reference the actual beta was most similar to by absolute difference. The CpG was then tallied to the most similar comparison group. To visualize the mixture reference, a 60:40 weighted average of the reference B-memory and B-naive CpGs was used, which was based on the mean proportions of the predicted mixed samples. B-Index To develop B-Index, the prediction score for CLL, where 0 corresponds to epigenetically B-naive-like, and 1 corresponds to epigenetically B-memory-like, techniques based on DNA methylation cell type deconvolution were used. The 12-cell reference library (described above) was subsetted to the B-memory and B-naive references, with the top-most differentially methylated CpGs between the two references available in our data included. Constrained projection was then performed using the projectCellType_CP function from the FlowSorted.BloodExtended.EPIC package (version 0.99.0)23. The two resulting proportions were restricted to a minimum of 0. The results were then normalized to sum to 1. The B-Index was then defined to be the B-memory proportion after the previously described operations. The equation for B-Index for a sample given predicted B-memory proportion p^Bmem and predicted B-naive proportion p^Bnv is shown below: To determine how many of the top-most differentially methylated CpGs between the B-naive and B-memory references to include, varying numbers of CpGs, from 10 up to the total number available in our data from the reference library, were used to calculate the B-Index for our samples. After visualization of the resulting B-Indexes on a spaghetti plot, 400 was selected due to the stabilization of B-Index variation per sample over increasing amounts of included CpGs. Epigenome-wide association studies Multiple epigenome-wide association studies (EWAS) were conducted. EWAS analyses are linear regression models of CpG beta-values or M-values with respect to input parameters. For every EWAS in this study, M-values, which are defined as log2β1−β, as described above, were used as the modeled variable in order to better fulfill the assumption of homoscedasticity, which beta value distributions tend to violate due to their bounding. For every EWAS, age and sex were adjusted for (except for when age and sex were unavailable in external cohorts). Age was centered around the mean value in the model matrix. The EWAS outputs linear models of the M-values, where beta coefficients of variables of interest and adjusting variables are estimated using linear regression. The resulting beta coefficient of the variable of interest (such as B-Index) is labeled as ΔM, which can be interpreted as the estimated change in M-value per unit increase of the variable of interest. Positive ΔM corresponds to hypermethylation at a given CpG with respect to the variable of interest, and negative ΔM corresponds to hypomethylation. For one EWAS, a transformation from ΔM to Δβ was conducted in order to apply a Δβ cutoff. To do this, the following equation was used (as described previously24), requiring the ΔM (beta-coefficient of variable of interest, when M-values used as output), and the intercept of the regression β0: EWAS were implemented using the limma package (version 3.62.2)25, using the lmFit function to perform regressions and the eBayes function to perform empirical Bayes moderation of standard errors. Adjusted p-values were calculated using the Benjamini–Hochberg false discovery rate method, with all inputted CpGs being used in the corrections. Enrichment analyses A number of enrichment studies were performed to assess whether a given list of CpGs has an overrepresentation of features of interest when compared to a background. Enrichment of Gene Ontology (GO) terms was performed to assess if CpGs of interest were enriched for functional gene groups. Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis was performed to assess enrichment for biological pathways. GO and KEGG analyses were performed using the gometh function in the MissMethyl package (version 1.40.3)26, using the 20a1.hg38 Illumina Epic V2 annotation (version 1.0.0)27. MissMethyl accounts for bias in both probes per gene on the EPICv2 array and CpG annotation to multiple genes, using Wallenius’ noncentral hypergeometric test. For CpG island context and chromosomal enrichment, the 20a1.hg38 Illumina Epic V2 annotation file was used. Two-sided Fisher’s exact test was used for chromosomal enrichment. For genomic context, annotation was constructed using AnnotationHub (version 3.14.0)28, and Ensembl (version 113). Genomic ranges of exons, introns, 5′UTR, 3′UTR, and promoters were accessed and mapped to probe coordinates using sesameData_getManifestGRanges. CpGs were marked as within promoter range up to 1500 bases upstream, and up to 200 bases downstream of promoters. Probes with multiple mappings were collapsed to a single genomic context using the following order of precedence: Promoter >5′UTR > 3′UTR > Exon > Intron. Probes not mapping to any genomic ranges were labeled intergenic. As CpG islands associate with genomic context, a Cochran-Mantel-Haenszel (CMH) test, stratifying across CpG island context, was used. Likewise, island context enrichment was tested using a CMH test, stratifying across Illumina probe type (Type I/Type II). For transcription factor binding site analysis, the KnowYourCG (KYCG) package (version 1.2.5)29,30 EPICv2 consensus transcription factor binding site (TBFS) knowledge base was used. KYCG consensus TBFS annotations are constructed by integrating ChIP-seq peaks from Cistrome and ReMap (derived from ENCODE), and ranking CpGs by overlap frequency among multiple datasets to capture consensus binding features. We collapsed this annotation to probe prefixes to match with our probe annotation, and removed the resulting duplicates. Enrichment tests were then conducted with KnowYourCG’s testEnrichment function, set to alternative=“greater”, which uses a one-sided Fisher’s exact test. Confidence intervals were added to the function output using the same statistical method. For all enrichment tests, a BH-adjusted FDR p-value cutoff of <0.05 was used. For all EWAS-derived enrichment analyses, all EWAS input CpGs were used as the total list, and CpGs were split into hypermethylated and hypomethylated for separate analyses. For the enrichment tests on CpGs determined to be similar to either B-naive or B-memory in U-CLL by TOST, all CpGs input into the TOST were used as the total list, in order to control for enrichment that B-cell-specific CpGs have in tested features, and enrichment analyses were performed separately on B-memory-like and B-naive-like CpGs. FDR p-values were calculated based on the number of tests conducted per analysis. Odds ratios with 95% confidence intervals were reported, as well as FDR BH-adjusted p-values. Forest plots are in log2 space. Normal B-cell references The 12 normal B-cell references (6 B-memory, 6 B-naive) were acquired via Gene Expression Omnibus (GEO) accession GSE174666. As described previously31, the B-memory and B-naive cells were isolated via magnetic activated cell sorting (MACS) techniques. Leukocytes from adult whole blood were separated into CD27+ and CD27−, and then CD19+ and CD19−. B-memory cell populations (CD19+ and CD27+), and B-naive cell populations (CD19+ and CD27−), were reported to have purities between 90% and 98%. Extracted DNA was measured on the Illumina EPICv1 microarray, and preprocessing was performed as described in the original study31. The processed beta matrix was retrieved from the GEO accession, along with phenotypic data. The beta values were converted to M-values, as described above. Equivalence testing To test equivalence between methylation states in U-CLL and the B-memory and B-naive references, we used the two one-sided t-tests (TOST) method, implemented in the TOSTER package (version 0.8.4)32. Among the results from the EWAS on the normal B-references, CpGs with FDR p < 0.05 and ∣Δβ∣ > 0.2 that were also available in our data were used for the analysis. At each of these CpGs, the U-CLL samples were tested for TOST equivalence to the B-naive and B-memory using an epsilon of 0.1, meaning a ±0.1 difference in beta value will be tolerated for equivalence. As standard with TOST analysis, the larger of the two p-values from the one-sided tests was taken as the p-value for the TOST. Benjamini–Hochberg FDR-adjusted p-values were calculated based on the total CpGs input to the TOST. Validation analyses To validate results from our internal cohort of 89 samples, two external, publicly available cohorts were used. The first was GSE13672433, which contains 72 samples measured on the 450k array. One sample, a later longitudinal sample from an already represented patient, was filtered out. Beta values were derived from the matrix deposited on GEO, and clinical subtype from the IGHV mutation information available. B-Index was predicted on this cohort by subsetting the 400 CpGs used for B-Index to the ones available in this data, which resulted in 140 CpGs. The second external cohort (GSE289415)34, contained 52 U-CLL samples on EPICV1 array. B-Index was predicted using all 400 CpGs. These U-CLL cases underwent the same course of TOST analyses described above, using all the B-memory and B-naive differentially methylated CpGs that were available on the array (97,407 CpGs out of 97,709), using the same parameters as described above. Additional statistical analyses Analyses were performed on R (version 4.4.1)35. Wilcoxon rank-sum test was performed using the wilcox.test function, without continuity correction. Receiver operating characteristic (ROC) curves and Youden’s J statistics were generated using the PRROC package (version 1.4)36. All box plots were made with default ggplot settings; the center line indicates the median, the box bounds indicate Q1 and Q3, and the whiskers extend to the most extreme values within the 1.5× IQR. Results Development of B-Index We analyzed residual peripheral blood samples from 89 patients with chronic lymphocytic leukemia that were collected during their routine laboratory workups (Table 1). The study population consisted of 56 IGHV mutated (63%), 24 IGHV unmutated (27%), 3 Polyclonal IGHV (3%), and 6 cases where IGHV status was unavailable (7%). Subjects were predominantly male (67%), and the average age was 71.3. Most subjects had not received treatment for CLL (66%). Information on IGHV mutational percentage was available for 46 participants. Table 1Cohort characteristicsCharacteristicOverall n = 89Value(%)Age at draw (years)71.3 ± 9.1SexMale60(67)Female29(33)IGHV mutation statusMutated56(63)Unmutated24(27)Polyclonal3(3)Failed6(7)Treated at the time of sample collectionYes30(34)No59(66)Clonal B-cell count at diagnosis (×103/μL)37.4 ± 68.0Sample lymphocyte count (×103/μL)40.1 ± 74.2Continuous variables: mean ± SD. Categorical variables: n and (%). Cohort characteristics Continuous variables: mean ± SD. Categorical variables: n and (%). CLL tumor cells were isolated from peripheral blood using a magnetic bead negative selection kit. DNA was extracted from isolated cells, quantified, and measured on high-density methylation microarrays. Quality control was performed on the methylation data, with all samples passing checks (Supplementary Fig. 1A–D). Twelve cell type deconvolution was performed using the Flow-Sorted Extended reference library to assess fractions of B-naive, B-memory, and non-B-cell (basophil, CD8 memory T, CD8 naive T, eosinophil, neutrophil, natural killer, CD4 naive T, T regulatory, CD4 memory T, and monocyte) fractions. The deconvolution results consisted of almost entirely B-memory (Mean 0.81, SD 0.19) and B-naive (Mean 0.17, SD 0.19) fractions, showing very little signal from non-B-cell fractions, with the largest being monocytes (Mean 0.018, SD 0.037) (Fig. 1A, B). B-Index development. a Stacked bar plot depicting results from 12 cell type epigenetic deconvolution, split into tertiles based on B-memory proportion. B-memory (Mean 0.81, SD 0.19) and B-naive (Mean 0.17, SD 0.19) show the largest proportions, with the largest non-B-cell proportion being monocytes (Mean 0.018, SD 0.037). b Box plots depicting results of 12 cell-type deconvolution (same data as 1a, n = 89 biologically independent samples). A dashed vertical line separates B-cells from non-B-cells. c Heatmap of the top 100 most differentially methylated CpGs between B-memory and B-naive in the reference library. The samples with the most mixed predictions, taken from and in the same order as the right-most tertile of (a) (n = 30), are shown next to the B-memory and B-naive references, as well as an in-silico 60:40 mixture of B-memory and B-naive cells. d Box plots quantifying the number of CpGs per sample most similar to either (i) pure B-naive reference, (ii) pure B-memory reference, (iii) a custom mixture based on each sample’s deconvolved proportions (n = 30 biologically independent samples). Pure references explain the majority of CpGs (Mean 83%, SD 8.4%) in all samples except one. e Spaghetti plot showing the predicted B-Index of each sample as the number of CpGs included in prediction increases. CpGs were added in decreasing order of the most differentially methylated between B-memory and B-naive within the reference. f Violin plot showing distribution of the B-Index among all 89 samples. As many of the samples had both B-naive and B-memory components, we sought to determine if the components represented either a single purified cell population that has individual epigenetic features of both B-memory and B-naive cells, or a mixture of two distinct cell-type populations. To determine this, we tested the tertile of samples with the most mixed proportions (B-memory < 0.78, n = 30), and observed the top 100 most differentially methylated CpGs within the reference library between B-naive and B-memory, alongside an in-silico 60:40 mixture of B-memory and B-naive references. The samples were not similar to a mixture of two distinct cell populations (Fig. 1C, Supplementary Fig. 2). To quantify this, for each sample, we created a custom mixed comparison group based on deconvolution-derived proportions and the reference library at the same 100 CpGs. For each sample in the most mixed tertile, we compared beta values of the sample to the custom mixture and the pure B-memory and B-naive references (Fig. 1D). Results showed that for all the tested samples, more CpGs corresponded to a pure reference (Mean 83%, SD 8.4%) than to their custom mixture (Mean 17%, SD 8.4%), except for the sample assigned the largest monocyte proportion. This confirms that in the context of CLL cases and these references, deconvolution results imply an intermediate state, notably different than the usual use of deconvolution to quantify mixture proportions. This result also demonstrates clonality at the tested CpGs, as has been noted as a feature of the CLL epigenome previously37. Having confirmed that the deconvolution results in this context constitute an intermediate state within a purified population, we sought to adapt the deconvolution framework to produce a continuous epigenetic surrogate of each sample’s B-cell differentiation state. To do this, non-B-cell references were removed from the 12-cell reference library, and the most differentially methylated CpGs between B-naive and B-memory were used for constrained projection. After performing constrained projection on a sample, the two resulting predictions of B-naive and B-memory were rescaled to sum to one. The B-Index is then defined as the predicted B-memory score after this rescaling. A score of 0 indicates high epigenetic similarity to B-naive, and a score of 1 indicates high similarity to B-memory. Samples can take on any continuous value across the 0-1 range. To determine how many CpGs to use for constrained projection, CpGs were added in order of the most differentially methylated (Fig. 1E). In total, 400 were selected as a cutoff due to the stabilization of B-Index prediction. Principal component analyses of the 89 CLL samples showed that B-Index varied with major components within the 400 CpGs used to develop B-Index, and across all CpGs measured in the epigenome (Supplementary Fig. 3). Among the full cohort, the distribution of B-Indexes was bimodal, with no sample having a B-Index lower than 0.28 (Fig. 1F). Validation and epigenetic correlates of B-Index B-Index values were significantly different (p = 2.6 × 10−12) between M-CLL (Mean 0.92, SD 0.094) and U-CLL (Mean 0.48, SD 0.096) subgroups by two-sided Wilcoxon rank-sum test (Fig. 2A). To evaluate the utility of B-Index for inferring IGHV mutation status, we used it to classify samples into U-CLL and M-CLL. The area under the curve of the Receiver Operating Characteristic (ROC) was 0.995 (Fig. 2B). An optimal B-Index cutoff within the 89-sample cohort was found to be 0.60 by maximizing Youden’s J statistic. Using that threshold, B-Index was able to correctly classify samples into U-CLL and M-CLL with an accuracy of 98.8%. Samples without definitive IGHV mutational status information were excluded from classification analysis (Polyclonal n = 3, Failed n = 6). As U-CLL and M-CLL are categorical distinctions derived from tumor IGHV percent identity to germline, we compared B-Index to this underlying continuous variable. We found that IGHV mutational percentage and B-Index were correlated among the 46 (29 M-CLL and 17 U-CLL) samples where IGHV percentage was available (Supplementary Fig. 4A), with an R2 of 0.65 (p = 1.1 × 10−11). Validation and epigenetic correlates of B-Index. a Box plot of B-Index stratified into IGHV Mutated (M-CLL) and IGHV Unmutated (U-CLL). Mean B-Index of M-CLL was 0.92 (SD 0.094), mean B-Index for U-CLL was 0.48 (SD 0.096), the two distributions were significantly different by two-sided Wilcoxon test (n = 80 biologically independent samples, p = 2.6 × 10−12). b Receiver operating characteristic (ROC) curve for B-Index as a classifier into U-CLL and M-CLL. AUC was 0.995, and classification had 98.8% accuracy at Youden’s J optimal cutoff of 0.6. c Volcano plot depicting results from epigenome-wide association study (EWAS) performed on B-Index. x-Axis is delta M, or the change in M-value per unit increase of the predictor (B-Index). The horizontal line marks the p-value corresponding to Benjamini–Hochberg (BH) FDR-adjusted p < 0.05. In total, 83,129 CpGs were significantly hypomethylated, and 30,710 were hypermethylated. d Forest plot of island context enrichment among the hypo and hypermethylated CpGs with increasing B-Index, from (c). Points indicate log2 odds ratios, and error bars indicate 95% confidence intervals. Entries in black are significant (FDR p < 0.05). e Forest plot of transcription factor binding site enrichment (one-sided Fisher’s exact test) among the hypo and hypermethylated CpGs with increasing B-Index, from (c). Points indicate log2 odds ratios, and error bars indicate 95% confidence intervals. Entries in black are significant (FDR p < 0.05). f Venn diagram of FDR-significant CpGs from an EWAS conducted on IGHV mutational percent and on B-Index within the subset where IGHV mutational percent was available. We also calculated B-Index on a cohort of 71 CLL samples (GSE136724), described previously33, measured on a different array type (Illumina 450k). We used 140 of the original 400 CpGs to calculate B-Index, as only this subset of CpG sites was measured on the 450k array. This external validation set retained a significant difference in B-Index across clinical subtype (p = 1.0 × 10−7), with U-CLL cases having a mean B-Index of 0.55 (SD 0.13), and M-CLL of 0.90 (SD 0.11) (Supplementary Fig. 5A). Using B-Index as classification on this group yielded an AUC of 0.975, and an accuracy of 91.5% with a cohort optimal B-Index cutoff of 0.68 (Supplementary Fig. 5B). We calculated B-Index for an additional external cohort of just U-CLL cases, measured on the EPICv1 array (GSE289415)34. All 400 B-Index CpGs were present in the data. B-Indexes fell within the expected range for U-CLL cases (mean 0.48, SD 0.08), with no cases with B-Indexes above 0.65 (Supplementary Fig. 5C). To determine the changes in the CLL methylome associated with B-Index, we used an epigenome-wide association study (EWAS) with B-Index as the outcome variable. Using the full internal cohort (n = 89) and adjusting for the well-documented epigenetic confounders of age and sex, increasing B-Index showed strong signals skewed toward hypomethylation, with 83,129 hypomethylated CpGs and 30,710 hypermethylated CpGs passing the false discovery rate (FDR) adjusted p-value threshold of 0.05 (Fig. 2C) (Supplementary Data 1). Among the genes related to the most significant hypomethylated CpGs were LRRC31 (FDR p = 3.6 × 10−30), ANO7 (FDR p = 1.1 × 10−28), and OGG1 (FDR p = 3.6 × 10−28). Among the genes related to the most significant hypermethylated CpGs were PRR33 (FDR p = 2.6 × 10−26) and PPCDC (FDR p = 3.7 × 10−25). Many of the most significant hypermethylated CpGs had multiple redundant gene hits, such as four CpGs related to CRY1 (FDR p-values: 7.7 × 10−23, 2.0 × 10−20, 2.7 × 10−20, 2.3 × 10−18), two related to LMAN2 (FDR p-values: 5.7 × 10−26, 2.4 × 10−23), and two related to CLDN15 (FDR p-values: 6.6 × 10−19, 9.8 × 10−19). Note that an EWAS on categorical treatment status at sample collection did not identify significant associations with methylation, and EWAS models therefore did not include adjustment for treatment status (Supplementary Fig. 6). To characterize the epigenetic signals correlated with B-Index at the hypomethylated and hypermethylated CpGs, we used enrichment analyses across multiple CpG attributes. We conducted a chromosomal enrichment analysis of the significantly (FDR p-value < 0.05) hypo and hypermethylated CpGs using Fisher’s exact test (Supplementary Fig. 4B, Supplementary Data 2 and 3). Among the hypomethylated CpGs, there was enrichment of chromosome 18 (OR = 1.20, FDR p = 2.7 × 10−12), 20 (OR = 1.15, FDR p = 3.9 × 10−10), and 21 (OR = 1.20, FDR p = 8.0 × 10−9) among others. Representation of chromosome 19 was significantly depleted (OR = 0.68, FDR p = 4.7 × 10−89), among others. The hypermethylated CpGs were enriched for chromosome 17 (OR = 1.27, FDR p = 6.2 × 10−21), among others. Chromosomes 2 (OR = 0.82, FDR p = 5.4 × 10−17) and 5 (OR = 0.80, FDR p = 4.7 × 10−15) were depleted, among others. We also performed an enrichment analysis on CpG island relation, within the same sets of hyper/hypomethylated CpGs, adjusting for Illumina probe type using a Cochran–Mantel–Haenszel (CMH) test (Fig. 2D, Supplementary Data 4 and 5). Hypomethylated CpGs were depleted in CpG islands (OR = 0.14, FDR p < 10−100) and shore regions (OR = 0.62, FDR p < 10−100), and enriched in open-sea regions (OR = 2.3, FDR p < 10−100). Hypermethylated CpGs were enriched for representation of CpG islands (OR = 1.14, FDR p = 1.5 × 10−13), and depleted in open-sea (OR = 0.93, FDR p = 6.5 × 10−9). We further conducted an enrichment analysis on CpG genomic context, adjusting for island relation using a CMH Test (Supplementary Fig. 4C, Supplementary Data 6-7). Hypomethylated CpGs were enriched in intergenic (OR = 1.3, FDR p < 10−100) and intron regions (OR = 1.16, FDR p = 1.2 × 10−82), and depleted in promoter (OR = 0.81, FDR p < 10−100), 5′UTR (OR = 0.60, FDR p = 1.6 × 10−49), 3′UTR (OR = 0.68, FDR p = 1.0 × 10−78), and exon regions (OR = 0.84, FDR p = 8.3 × 10−38). Hypermethylated regions were enriched for 5′UTR (OR = 1.24, FDR p = 6.3 × 10−11) and intron regions (OR = 1.29, FDR p = 3.3 × 10−93), and depleted in promoter (OR = 0.82, FDR p = 1.7 × 10−42) and intergenic regions (OR = 0.69, FDR p = 4.5 × 10−60). As transcription factor binding sites (TFBS) are known to change methylation states throughout many maturation pathways, we checked for enrichment of CpGs near TFBS within our CpGs that we found to be hypo/hypermethylated with increasing B-Index (Fig. 2E, Supplementary Data 8 and 9). Hypomethylated CpGs were enriched most significantly for binding sites of BATF3 (OR = 3.25, FDR p = 1.1 × 10−76). Hypermethylated CpGs were enriched in binding sites of JARID2 (OR = 1.99), EZH2 (OR = 2.02), LMO2 (OR = 2.12), and LYL1 (OR = 2.32) (all listed FDR p < 10−100), among others. We also assessed enrichment among gene ontology (GO) terms. Hypomethylated CpGs showed enrichment for cell periphery (FDR p = 1.5 × 10−43), plasma membrane (FDR p = 4.0 × 10−35), and signaling receptor activity terms (FDR p = 2.0 × 10−16), among others (Supplementary Fig. 4D, Supplementary Data 10). An enrichment analysis of Kyoto Encyclopedia of Genes and Genomes (KEGG) terms showed significantly hypomethylated pathways in cornified envelope formation (FDR p = 1.72 × 10−11), among others (Supplementary Fig. 4E, Supplementary Data 11). For FDR significantly hypermethylated CpGs, KEGG analysis showed no significant results after FDR correction (Supplementary Data 12). GO analysis on the hypermethylated CpGs showed some FDR-significant results, such as regulation of developmental processes (FDR p = 3.9 × 10−3) (Supplementary Data 13). To further assess the strength of the epigenetic correlates of B-Index in comparison to the underlying continuous clinical parameter, IGHV mutational status, we performed EWAS analyses against B-Index and IGHV percentage germline identity on the subset of data where IGHV percent germline identity was available (n = 46, 29 M-CLL and 17 U-CLL cases), adjusting for age and sex (Supplementary Fig. 7A, B, Supplementary Data 14 and 15). The B-Index EWAS identified 14,776 significantly (FDR p < 0.05) hypermethylated CpGs, and 48,982 significantly (FDR p < 0.05) hypomethylated CpGs. IGHV mutational percent was significantly (FDR p < 0.05) associated with fewer CpGs than B-Index: 12,218 hypermethylated and 32,250 hypomethylated. The B-Index differentially methylated CpGs largely encompassed those found by the IGHV mutational percent EWAS (Fig. 2F). To determine whether continuous B-Index is able to identify additional epigenetic information within each subgroup, we stratified our cohort into U-CLL (n = 24) and M-CLL (n = 56) cases and performed an EWAS for B-Index on each. The EWAS on the U-CLL cases (Supplementary Fig. 7C) found no FDR p < 0.05 significant hits. The EWAS on the M-CLL cases found 2,221 FDR p < 0.05 significantly hypermethylated CpGs, and 16,459 significantly hypomethylated (Supplementary Fig. 7D, Supplementary Data 16). A Venn diagram of the B-Index-associated CpGs from M-CLL and the B-Index-associated CpGs from the full CLL cohort showed a partial overlap of the two signals (Supplementary Fig. 7E). As no significant CpGs were found to be associated with B-Index in our 24 CLL samples, we assessed if an association with B-Index could be detected in a larger (n = 52) cohort of U-CLL cases. The EWAS (unadjusted for age and sex due to lack of available covariate information), detected 1,653 significantly (FDR p < 0.05) hypermethylated CpGs, and 20,229 significantly hypomethylated CpGs, confirming epigenetic associations with B-Index in both clinical subgroups (Supplementary Fig. 5D, Supplementary Data 17). U-CLL similarities to normal B-cell subsets The U-CLL cases were found to have a mean B-Index of 0.48, indicating some commonalities with both B-naive and B-memory cells. We subsequently sought to further investigate the epigenetic similarities of U-CLL to B-naive and B-memory cells in terms of normal biology rather than CLL pathology. To serve as non-pathologic B-cell epigenome references, we used high-density methylation microarray data (Illumina EPICv1 platform) consisting of 6 purified B-naive cell samples and 6 B-memory cell samples from healthy donors. Each cell population group was evenly split between male and female donors, with average ages of 38 (SD 10) for the B-naive donors and 32 (SD 12) for the B-memory donors. We first determined the differentially methylated CpGs across the B-naive and B-memory cell populations, adjusting for the age and sex of the donors (Supplementary Fig. 8, Supplementary Data 18). After imposing a cutoff of FDR-adjusted p < 0.05, and a minimum change in beta value of 0.2, there were 108,978 significantly hypomethylated CpGs in B-memory compared to B-naive, and 11,156 significantly hypermethylated. Among the 120,134 differentially methylated CpGs, 97,709 were represented in the EPICv2 array and available in our CLL dataset. Visualizing CLL cases’ methylation at the top 1000 most significant differentially methylated CpGs between B-naive and B-memory cells showed strong similarity between U-CLL and B-memory cells (Fig. 3A). In contrast, plotting the top 1000 most significantly differentially methylated CpGs from an EWAS comparing U-CLL to M-CLL (Supplementary Fig. 9A, Supplementary Data 19) showed the opposite pattern (Fig. 3B), with the U-CLL cases appearing similar to B-naive cells. U-CLL and M-CLL are thus both more memory-like than naive-like under the normal reference-derived CpGs, and polarize to the references when using the CpGs differentially methylated between U-CLL and M-CLL. Heatmaps of CLL methylation alongside B-memory and B-naive cells, across different CpG sets. a Heatmap of the top 1000 most significant CpGs from the B-memory v. B-naive EWAS available in our data. B-memory and B-naive references are shown alongside all samples, with IGHV mutational status shown in the top bar, and sorted in order of increasing B-Index from left to right. b Heatmap similar to (a), except using the top 1000 most significant CpGs from the U-CLL vs M-CLL EWAS. To quantify U-CLL’s epigenetic similarity to B-memory and B-naive cells, we performed equivalence tests between U-CLL and the normal B-cell groups on the 97,709 differentially methylated CpGs between B-memory and B-naive cells available in our data. We used the two-one-sided-T (TOST) test with an equivalence range of beta ±0.1. Among the beta values of CpGs tested, 6046 were significantly (FDR p < 0.05) equivalent in U-CLL to B-memory cells, and 522 to B-naive cells (Fig. 4A, Supplementary Data 20). To verify that the B-memory-like signature of U-CLL was not contingent on our CpG prefiltering, we computed the Manhattan distance of each U-CLL case to the mean beta values among the B-memory and B-naive cases among all the 682,014 CpGs available across both datasets (Supplementary Fig. 9B). Every U-CLL case had a higher Manhattan distance from B-naive than B-memory, indicating higher relative similarity with B-memory than B-naive. U-CLL similarities to B-memory and B-naive cells. a Bar graph depicting results of two one-sided t-tests (TOST) with equivalence bounds of ±0.1, assessing which CpGs in U-CLL are FDR p < 0.05 significantly equivalent to either B-memory or B-naive among the BH FDR-significant and differentially methylated CpGs across B-memory and B-naive cells (see Supplementary Fig. 8). b Contingency table showing the methylation state of the TOST FDR p < 0.05 CpGs based on the normal B-cell to which they were equivalent in U-CLL, and their methylation state in U-CLL. c Forest plot of transcription factor binding site enrichment among CpGs in U-CLL, similar (TOST FDR p < 0.05) to B-memory and B-naive. Points indicate log2 odds ratios, and error bars indicate 95% confidence intervals. Entries in black are significant (one-sided Fisher’s exact test, BH-corrected FDR p < 0.05). To further characterize the B-naive-like and B-memory-like signatures of U-CLL, we separated the FDR p < 0.05 CpGs from the TOST analysis into those that are in the hypermethylated or hypomethylated state in U-CLL, defined based on the direction of differential methylation in the B-naive and B-memory EWAS (Fig. 4B). Results showed that the U-CLL’s B-memory-like signature is hypomethylated, and its B-naive-like signature is mostly hypermethylated, with a chi-squared test yielding a highly significant association (−log10(p)>500). In order to better discern in what particular ways U-CLL is B-memory and B-naive-like, we conducted a series of enrichment analyses on both groups of the TOST equivalent CpGs. We used the TOST CpGs with FDR p < 0.05, and the 97,709 CpGs found to be different between B-memory and B-naive cells as the background to remove B-cell-specific enrichment that is in common between B-memory and B-naive. Results for chromosomal enrichment were insignificant at all chromosomes for B-naive-like CpGs, and showed depletion at chromosome 5 in B-memory-like CpGs (OR = 0.80, FDR p = 5.1 × 10−3) (Supplementary Data 21 and 22). CpG island relation enrichment was assessed using a CMH test, adjusting for Illumina probe type (Supplementary Fig. 9C, Supplementary Data 23 and 24). B-memory-like CpGs were enriched for shore regions (OR = 1.25, FDR p = 3.5 × 10−9), and depleted in island (OR = 0.56, FDR p = 1.9 × 10−8) and open-sea (OR = 0.93, FDR p = 1.7 × 10−2) regions. B-naive-like CpGs had no detectable enrichment signal. We then conducted enrichment analysis on genomic context, adjusting for island relation using a CMH Test (Supplementary Fig. 9D, Supplementary Data 25 and 26). B-memory-like CpGs in U-CLL showed depletion in 3′UTR (OR = 0.71, FDR p = 6.8 × 10−4). B-naive-like CpGs showed significant enrichment in 3′UTR (OR = 1.75, FDR p = 2.2 × 10−2) and exons (OR = 1.38, FDR p = 4.8 × 10−2), and depletion in intergenic regions (OR = 0.72, FDR p = 4.8 × 10−2). Enrichment of representation of GO and KEGG terms was also assessed among the B-memory and B-naive-like CpGs, returning no FDR p < 0.05 significant results (Supplementary Data 27–30). As transcription factor binding site (TFBS) methylation is known to change at many points along the normal B-naive to B-memory development, we assessed enrichment for TFBS among the B-memory and B-naive-like CpGs in U-CLL (Fig. 4C, Supplementary Data 31 and 32). Among the many notable TFBS we found to be enriched among the CpGs similar to B-memory cells were POU2F2 (OR = 6.53), TBX21 (OR = 4.85), RELA (OR = 5.73), IRF4 (OR = 4.08), PAX5 (OR = 4.98), and BATF (OR = 4.15) (all listed FDR p < 10−200) among others. Outside of the top twenty shown in the figure were additional FDR-significant TFBS of interest, such as the NF-κB family members: NFKB1 (OR = 3.54), NFKB2 (OR = 3.76), REL (OR = 5.90), and RELB (OR = 4.02), as well as BCL6 (OR = 7.25) and MEF2C (OR=6.45) (all listed FDR p < 10−100). Enriched binding sites among the CpGs similar to B-naive included LMO2 (OR = 5.03, FDR p = 3.3 × 10−14), RUNX1 (OR = 5.33, FDR p = 2.1 × 10−11), and FLI1 (OR = 6.66, FDR p = 2.0 × 10−10), among others. To validate the similarity of U-CLL to B-memory and B-naive cells, we conducted additional analyses on an external cohort of 52 U-CLL samples (GSE289415)34. The TOST test was repeated on this cohort for B-memory and B-naive similarity at the 97,407 differentially methylated CpGs between B-memory and B-naive cells available in the external cohort data. Similar to our data, a dominant B-memory-like signature was identified in this cohort with 13,596 CpGs being significantly (FDR p < 0.05) equivalent in methylation to B-memory cells, and 760 being significantly equivalent to B-naive cells (Supplementary Fig. 10A, Supplementary Data 33). Most CpGs identified as B-memory-like from the initial cohort (among those available in both cohorts) were also significantly B-memory-like in the validation cohort (87.8%), and similarly for B-naive-like CpGs (51.6%) (Supplementary Fig. 10B). CpG sets identified in the initial and validation cohorts had highly significant overlap for both B-naive-like CpGs (p < 10−300), and B-memory-like CpGs (p < < 10−300), indicating consistency in methylation patterns across cohorts. Epigenetic correlates of disease burden In addition to IGHV mutational data, for each study participant we also gathered two metrics of disease burden: the patient’s clonal B-cell count (×103 cells μL−1) at the time of their diagnosis (Mean 37 SD 68), and the total lymphocyte count (×103 cells μL−1) in the sample contemporary with their epigenetic data (Mean 40 SD 74) (Supplementary Fig. 11A). The two metrics, in log2 space, had R2 of 0.53 (Supplementary Fig. 11B). To assess the epigenetic correlates of each disease burden metric, EWAS analyses were performed on total lymphocyte count (Fig. 5A, Supplementary Data 34) and clonal B-cell count at diagnosis (Supplementary Fig. 11C, Supplementary Data 35) both in log2 space, adjusting for age and sex. 20,973 CpGs were found to be significantly (FDR p-value < 0.05) hypomethylated with increasing total lymphocyte count, and 2484 were significantly hypermethylated. In line with the other burden metric, clonal B-cell count showed a hypomethylation-skewed epigenetic signature, with 16,475 significantly hypomethylated CpGs and 997 hypermethylated. Based on the total lymphocyte count’s more significant epigenetic signal, it was used as the tumor burden metric of choice in all downstream analysis. To ensure this burden metric was independent from B-Index, a Pearson correlation test was performed across all samples on the log2 transformed burden metric, which showed no significant linear relationship (R2 = 0.004, p = 0.57) (Fig. 5B). Adjusting for B-Index in the total lymphocyte count EWAS also showed little change in significant CpGs (Supplementary Fig. 11D, Supplementary Data 36). Epigenetic correlates of tumor burden. a EWAS on total lymphocyte count in the sample, conducted in log2 space. 20,973 CpGs were significantly (BH corrected FDR p < 0.05) hypomethylated with increasing tumor burden, and 2484 were significantly hypermethylated (8.4-fold more significantly hypomethylated sites than hypermethylated). b Scatterplot of log2 total lymphocyte count and B-Index among the samples (n = 89, R2 = 0.004, p = 0.57). Colors of dots depict IGHV-mutational status. The line of best fit is shown. c Top 4 hits from KEGG pathway enriched among hypermethylated and hypomethylated CpGs with increasing tumor burden, analyzed separately. Enrichment was assessed using Wallenius’ noncentral hypergeometric test, and unadjusted and BH FDR-adjusted p-values are shown. To characterize the hypomethylation and hypermethylation signatures, we performed a series of enrichment analyses, beginning with chromosomal representation (Supplementary Fig. 12A, Supplementary Data 37 and 38). Hypermethylated CpGs were enriched for chromosome 13 (OR = 1.63, FDR p = 2.1 × 10−4), and depleted in chromosome 18 (OR = 0.61, FDR p = 2.7 × 10−2), among other significant deviations. Hypomethylated CpGs were enriched for chromosome 13 (OR = 1.29, FDR p = 3.6 × 10−8) and depleted in chromosome 7 (OR = 0.84, FDR p = 5.3 × 10−7), among others. We also assessed enrichment of CpG island context, adjusting for Illumina probe types using a CMH test (Supplementary Fig. 12B, Supplementary Data 39 and 40). Hypomethylated CpGs were enriched for open-sea regions (OR = 1.22, FDR p = 1.5 × 10−39) and were depleted in island regions (OR = 0.59, FDR p = 8.5 × 10−88). Hypermethylated CpGs were enriched in shore (OR = 1.40, FDR p = 9.4 × 10−12) and shelf (OR = 1.44, FDR p = 1.3 × 10−6) regions, and depleted in open-sea (OR = 0.86, FDR p = 3.5 × 10−4) and island (OR = 0.77, FDR p = 8.7 × 10−8) regions. We then analyzed enrichment of genomic contexts using CMH tests, adjusting for CpGs relationship to island (Supplementary Fig. 12C, Supplementary Data 41 and 42). No significant relationship was detected in the hypermethylated CpGs. The hypomethylated CpGs were enriched for intron regions (OR = 1.16, FDR p = 4.6 × 10−24), and depleted in intergenic (OR = 0.78, FDR p = 8.5 × 10−23), among others. We then assessed enrichment of transcription factor binding sites (Supplementary Fig. 12D, Supplementary Data 43 and 44). Hypomethylated CpGs were enriched for many sites of interest, including RUNX3 (OR = 3.04), EBF1 (OR = 3.29), BATF (OR = 3.65), EBNA3C (OR = 4.98) (all listed FDR p < 10−300), among others. Notably, EBNA3 (OR = 7.70, FDR p < 10−100), and EBNA2 (OR = 4.10, FDR p = 1.3 × 10−67), were also enriched. Hypermethylated regions were enriched in binding sites for EZH1 (OR = 2.06, FDR p = 1.9 × 10−23), UHRF2 (OR = 2.05, FDR p = 2.3 × 10−21), SOX4 (OR = 1.89, FDR p = 3.1 × 10−19), among others. We further performed enrichment analyses on gene ontology (GO) terms and KEGG pathways (Supplementary Data 45–48). Among the significantly hypomethylated with increasing tumor burden, GO enrichment analysis revealed FDR-significant hits for cytosol (FDR p = 4.1 × 10−9) and lymphocyte activation (FDR p = 5.2 × 10−9), among others (Supplementary Fig. 12E). KEGG pathway analysis on hypomethylated CpGs revealed enrichment for the B-cell receptor signaling pathway (FDR p = 9.2 × 10−4), Epstein–Barr virus infection (FDR p = 2.9 × 10−4), herpes simplex virus 1 infection (FDR p = 5.1 × 10−4) and viral carcinogenesis (FDR p = 1.2 × 10−3), among others (Fig. 5C). Hypermethylated CpGs showed KEGG pathway enrichment of neutrophil extracellular trap formation (FDR p = 2.4 × 10−5) among others, and GO hits in structural constituent of chromatin (FDR p = 1.8 × 10−4), among others. Discussion We developed a continuous methylation-based metric for the B-differentiation state of CLL (B-Index) by applying epigenetic cell type deconvolution methods to a clonal population. B-Index was able to discriminate between U-CLL and M-CLL with high accuracy. We also show that B-Index carries additional epigenetic correlates within the M-CLL and U-CLL subgroups, demonstrating the informative value that a continuous metric can provide, as noted previously14. B-Index carried a stronger epigenetic signal than the continuous metric underlying the clinical subgroup designation, IGHV mutational status. B-Index was validated on multiple external cohorts, providing discrimination between U-CLL and M-CLL, and predicting consistent B-Indexes within U-CLL groups. Higher B-Index was correlated mostly with hypomethylation, largely preferential to open-sea regions, and rarely occurring at CpG islands. Broad hypomethylation has been previously noted as characteristic of both B-cell differentiation and CLL pathology13,38. B-Index was associated with methylation differences across CLL samples near some TF binding sites. Hypomethylation near BATF3 binding sites was associated with increasing B-Index. BATF3 has been previously implicated in Anaplastic Large Cell Lymphoma as being overexpressed and essential to tumor growth39. Hypomethylation at BATF3 binding sites may increase chromatin accessibility or transcription factor binding ability40. Hypermethylation with increasing B-Index occurred at sites near EZH2 binding, which is the catalytic subunit in the Polychrome Repressor Complex 2 (PRC2)41, and JARID2, which is known to recruit the PRC242. PRC2 activity has been shown to be up-regulated in other hematological malignancies, such as B-cell lymphomas43. EZH2 has also been implicated in lymphopoiesis and immunoglobulin heavy chain rearrangement44. Hypermethylation also occurred at sites near LMO2 and LYL1 binding, both of which are known to play roles in lymphocyte differentiation45,46, and have been shown to have interactions essential to key features of Acute Lymphoblastic Leukemia (ALL) in mice47. Importantly for interpretation, methylation at CpG sites associated with B-Index is necessarily variable across CLL cases, meaning that these sites can be both hypermethylated or hypomethylated in CLL neoplasms, in a manner correlated to the B-Index. In previous literature, U-CLL has often been described as epigenetically B-naive-like10,13,48, however, other previous studies have found the opposite characterization14. In this study, we are able to replicate both findings, depending on the approach to data analysis. We find that when a common prefiltering of CpGs that are most differentially methylated between U-CLL and M-CLL is applied, U-CLL appears very B-naive-like. However, we also find that when prefiltering is performed in terms of normal tissue biology by selecting the CpGs most differentially methylated between B-naive and B-memory cells, U-CLL is strongly B-memory-like, with ~12-fold more CpGs similar to B-memory cells than B-naive cells. This result remained consistent on an external cohort of 52 CLL cases. We also find that when no prefiltering is applied, every one of our U-CLL cases remains more epigenetically similar to B-memory than B-naive. The origin of the different findings is that prefiltering CpGs to those most different or variable between or within a group of M-CLLs and U-CLLs removes their common epigenetic signal, which is B-memory-like. We further identified the CpGs in U-CLL that are similar in methylation to B-memory cells, as well as B-naive cells. We found that CpGs with a B-memory-like methylation state in U-CLL are overwhelmingly hypomethylated, and the B-naive-like CpGs in U-CLL are largely hypermethylated. As B-cell differentiation consists of multiple epigenetic steps with a strong bias toward hypomethylation38, our results are consistent with U-CLL adopting most of the B-memory epigenetic transition. In our transcription factor binding site analysis of B-memory similar CpGs in U-CLL, we found enrichment for many binding sites of TFs related to the germinal center response pathway49, such as BCL6, MEF2C, IRF4, POU2F2 (OCT2), and NF-κB family transcription factors (NFKB1, NFKB2, REL, RELB). We also demonstrated that these B-memory-like CpGs were overwhelmingly hypomethylated. Many of the B-memory-like binding sites identified to be B-memory-like in U-CLL in this study were shown previously to become hypomethylated between the states of naive and germinal center states of B-cells, and largely retain that hypomethylation through their development into B-memory cells (IRF4, MEF2C, MTA3, BATF, PAX5) (see TF enrichment on their M4 CpG module)38. Our analyses provide evidence for further defining the cell type of origin and the initiation mechanism for U-CLL. As we have demonstrated that U-CLL has many aspects of a B-memory-like signature, models of cell of origin should account for this in context with other known facts about U-CLL, such as its low levels of somatic hypermutation, and reconcile these through a common mechanism. Two possibilities include that the cell of origin of U-CLL has a largely B-memory-like methylation profile when it begins its transformation to a neoplasm, or U-CLL gains this methylation profile after becoming neoplastic. As we demonstrated in this study, a continuum across B-differentiation states can provide additional epigenetic information outside of the usual classification, which could indicate multiple underlying substates of origin among the already-used subtypes of CLL. Future studies correlating B-Index to transcriptomic substates of B-cell differentiation would have the potential to provide additional biological verification of B-Index as a strong descriptor of CLL biology. We also investigated epigenetic associations with tumor burden in CLL. In our samples, tumor burden was independent of B-Index and IGHV mutational status, and had an associated methylation signature. Higher tumor burden was correlated with hypomethylation at roughly ~eightfold more CpG sites than hypermethylation. Hypomethylation was enriched in open-sea regions and depleted in islands. The hypomethylated epigenetic signal was enriched for a number of transcription factor binding sites, including MTA2 and MTA3 which have been shown to have relationships to metastasis and tumor growth in multiple cancers50–52. Multiple Epstein-Barr-Virus (EBV) transcription factor binding sites were enriched among the hypomethylated CpGs as well, including EBNA3C, EBNA3, and EBNA2. KEGG terms were also enriched for CpGs related to EBV infection, as well as Herpes simplex virus 1 infection, and viral carcinogenesis. EBV has been shown to have an impact on time to treatment and overall survival in CLL53, though the full relationship between EBV and CLL initiation and progression remains unclear54. Our data may provide motivation to analyze associations between the presence of EBV in CLL cells and tumor growth and burden metrics for other researchers. As our study used negative magnetic selection to enrich B-cells, additional research may be necessary to validate tumor burden epigenetic signatures in stronger purification environments, such as in CLL cells isolated via fluorescence-activated cell sorting (FACS). It may be the case that clonal B-cell count at the time of sampling would serve as a stronger epigenetic correlate than the total lymphocyte metric used in the study. As for the burden-correlated signals as a whole, with our study design, we are unable to parse whether the epigenetic signal is variable over time and acquired as the CLL cell population grows, or precedes and predicts CLL clonal population growth. Alongside our findings specific to CLL, we demonstrated that DNA methylation-based cell-type deconvolution methods can be used for constrained projection of intermediate cell states onto the purified cell population references, rather than solely for quantifying mixture compositions. This approach may be useful in developing continuous metrics for other malignancies that can originate at multiple points along normal differentiation pathways, though this approach is not necessarily limited strictly to references on the same pathway. A reasonable skeleton for determining if this method is useful in another pathology would be to perform initial multi-cell-type deconvolution on a purified malignant population of interest. If the results yield a small number of dominating proportions, those can be further examined to develop a continuous index, as in this study. Supplementary information Transparent Peer Review file Supplemental Information Description of Additional Supplementary files Supplementary Data 1-17 Supplementary Data 18-33 Supplementary Data 34-48 Supplementary Data 49 Supplementary Data 50 Supplementary Data 51 Supplementary Data 52 Supplementary Data 53 Acknowledgements The authors acknowledge the support of the Clinical Genomics and Advanced Technology Section in the Department of Pathology and Laboratory Medicine of the Dartmouth Hitchcock Health System, and the Pathology Shared Resource (RRID: SCR_023479) at the Dartmouth Cancer Center with NCI Cancer Center Support Grant 5P30 CA023108-41. Methylation microarray measurement was carried out in the Genomics and Molecular Biology Shared Resource (RRID:SCR_021293) at Dartmouth, which is supported by NCI Cancer Center Support Grant 5P30CA023108 and NIH S10 (1S10OD030242) awards. The authors also acknowledge pilot funding from the Dartmouth Cancer Center. Author contributions J.H. conceived of the study question, designed and carried out the analysis plan, and wrote the manuscript with input from all authors. B.C. assisted in the refinement of the analysis plan. B.C. and P.K. supervised the project, designed the initial study purpose, and acquired funding. L.W. conducted sample collection, sample processing, assisted in grant writing, and acquired clinical information. D.S. assisted in acquiring clinical information. S.S. assisted in sample processing. F.K.IV conducted methylation data collection. F.K. and P.P. contributed to data processing. All authors provided feedback throughout the study and approved the final version. Peer review Peer review information Communications Medicine thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available. Data availability Epigenetic and phenotypic data from the 89 samples used in this study, including raw IDAT files, the preprocessed beta matrix, B-Index, IGHV mutational status, IGHV mutational percentage, absolute clonal B-cell count at diagnosis, total lymphocyte count in the sample, age, and sex, are deposited on Gene Expression Omnibus (GEO) at accession GSE307998. Ages at 90 or above are bucketed and marked on GEO. Epigenetic and phenotypic data for the normal B-naive and B-memory references are available at Gene Expression Omnibus at accession GSE17466631. The 12-cell deconvolution reference data used to create the B-Index must be accessed through the original authors of the reference, available with a free license for academic use, described previously18. Source data for Figs. 1–5 are provided in Supplementary Data 49–53, respectively. Code availability All code used in analyses and figure generation is available on Zenodo55 at 10.5281/zenodo.17118029. Competing interests Dr. Christensen is a co-founder of Cellintec, which had no role in this work; the remaining authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Prabhjot Kaur, Brock Christensen. Supplementary information The online version contains supplementary material available at 10.1038/s43856-026-01666-8. References Associated Data Supplementary Materials Transparent Peer Review file Supplemental Information Description of Additional Supplementary files Supplementary Data 1-17 Supplementary Data 18-33 Supplementary Data 34-48 Supplementary Data 49 Supplementary Data 50 Supplementary Data 51 Supplementary Data 52 Supplementary Data 53 Data Availability Statement Epigenetic and phenotypic data from the 89 samples used in this study, including raw IDAT files, the preprocessed beta matrix, B-Index, IGHV mutational status, IGHV mutational percentage, absolute clonal B-cell count at diagnosis, total lymphocyte count in the sample, age, and sex, are deposited on Gene Expression Omnibus (GEO) at accession GSE307998. Ages at 90 or above are bucketed and marked on GEO. Epigenetic and phenotypic data for the normal B-naive and B-memory references are available at Gene Expression Omnibus at accession GSE17466631. The 12-cell deconvolution reference data used to create the B-Index must be accessed through the original authors of the reference, available with a free license for academic use, described previously18. Source data for Figs. 1–5 are provided in Supplementary Data 49–53, respectively. All code used in analyses and figure generation is available on Zenodo55 at 10.5281/zenodo.17118029.